Culture Japanese media catches wind of NovelAI Image Generation - People are pissed

Source: https://www.itmedia.co.jp/news/articles/2210/03/news146.html
Archive: https://archive.ph/Dnt4h

「とんでもなくハイクオリティー」 話題の画像AI「Novel AI」でひたすら二次元美少女と美少年を生成してみた
ENGLISH TRANSLATION FOLLOWS BELOW

Novel AI," a popular image AI, generates beautiful two-dimensional girls and boys.

NovelAI Diffusion," an image generation AI that began offering its service on October 3, has been attracting attention. Despite the fact that only paying members can use the service, it has already become a hot topic on Twitter, with users saying that it is "good at two-dimensional beautiful girls," and that "it comes up with a lot of outrageously high quality images without missing a beat in any way. As of 5 p.m., "Novel AI" was trending in Japan.

When the reporter actually joined the service and tried it out, he felt that, at least in the case of outputting two-dimensional beautiful girls and boys, he could obtain illustrations that were easier and of higher quality than any of the earlier image-generating AIs he had experienced. This article presents the images output by the reporter. I will examine the quality of the illustrations this service can produce and its strengths and weaknesses.

However, as discussed below, this AI tweeted that it was trained based on an image collection site suspected of violating copyright laws. This article uses the AI for the purpose of technical verification, but please keep this in mind when using the AI.

First of all, please take a look at this image. This is the image that came up as a result of typing "kawaii anime girl. The reporter recently wrote an article comparing various image-generating AI services. However, I felt that Novel AI was better at creating two-dimensional characters than any of the other services I used.
dy_pp_01.png

Let's try other keywords. I want to output a beautiful girl with blond hair and a red dress holding a sword, so I type "gold hair,red dress,anime girl having sword". This time, a beautiful girl with a bit of a sense of deja vu comes up. I was surprised that the so-called "air intake" (a hair style with two tufts of space on the top of the head) was also represented.
dy_pp_02.png

There is also a "variation" function that allows you to further arrange the resulting image, which I also used. The general atmosphere remains the same, but the details are changed.
1664823377433.png

Next, let's change things up a bit and specify a state other than appearance. The keywords are "sweat, cap, black long hair, manga girl. The next step was to specify a state other than appearance. Here is the result. It looks like she is sweating.
dy_pp_03.png


Twitter screenshots:
1664824201550.png1664824235049.png
1664824251882.png1664824270930.png1664824326310.png
 
Last edited:
Wait, japs are mad at AI created waifus? Is this one of those things where people strive to achieve something for years then the second it's within their grasp they begin hating it?
 
given that just about every hentai artist is interchangeable the fact that an AI can return higher quality art with simple text input should have all of them shaking in their fucking boots. imagine every one of those anime gallery sites being replaced with one site hosting a bot like this, running high power/availability network infrastructure. extend that to furry smut and like 90% of Patreon art whores would lose all their subscribers overnight

I'm also imagining a future where the basic bitch Patreon retards who read one HOW 2 DRAW HENTAI 4 DUMMIES book and started trying to monetize it being replaced by people making awful comics using AI-generated pictures cause the bots can't generate F-tier level hentai dialogue yet. maybe that's the next frontier
 
Wait, japs are mad at AI created waifus? Is this one of those things where people strive to achieve something for years then the second it's within their grasp they begin hating it?
Why would you want AI generated art anyway? Why not just replace beef patties with Impossible Whoppers?
 
Wait, japs are mad at AI created waifus? Is this one of those things where people strive to achieve something for years then the second it's within their grasp they begin hating it?
They have a very strange culture. This is a place where editing game saves is a highly illegal activity.

I can confirm that people are also training Stable Diffusion models on the corpus of images on Derpibooru and using it to generate ponies. :story:
Damn! Those are pretty good!
...and look! No wings! :heart-full::smug:
 
RoboChads keep on winning! Here's Craiyon from a few months ago:
View attachment 3712348View attachment 3712343
Eyes were definitely a struggle
I've noticed in AI generated art of different styles it'll sometimes fuck up the eyes and sometimes mouth to the point of something that looks like a rotten corpse face. Why does it have a more difficult time with facial features?
 
The answer is simple, if this sort of art can be replicated, artists will just draw things that for either ethical or practical reasons, cannot be by an AI.

Forces artists to develop a distinctive style now, which is good. Though I can hear the cries of a million coomer artists who have just been replaced by AI since all their style is generic garbage.

I've noticed in AI generated art of different styles it'll sometimes fuck up the eyes and sometimes mouth to the point of something that looks like a rotten corpse face. Why does it have a more difficult time with facial features?

@Mr. Bung

If I had to guess it's because all the expression is in the eyes and mouth. A nose is a nose but the eyes can be happy, sad, surprised, etc. Same for the mouth, open, closed, tense, pursed, smirk, etc. Unless it's specified then it's coming different looks and giving you the output. That's my guess I could be wrong.
 
I've noticed in AI generated art of different styles it'll sometimes fuck up the eyes and sometimes mouth to the point of something that looks like a rotten corpse face. Why does it have a more difficult time with facial features?
Understand my understanding of technology is... very rudimentary: essentially this is a philosophical question on what does "looks like" really mean? To us we have an established schema of what a person is and what eyes or a mouth look like. Even very slight variations from the norm are readily detected, even if someone cannot name them. That's not how a computer "thinks" though. It can only take very objective measures to then emulate. So in terms of hue/shape/etc it probably is just as accurate as the rest.
 
Back